Flow-Based Video Object Detection

Object detection is to identify the location and category of objects from captured images or videos through algorithms.As one of the most basic problems in computer vision area,object detection has important research value and broad application value.Video object detection requires accurately detect the location and category of objects in each video frame in real time.Existing works for video object detection require manually designed rules,and face huge challenges in terms of speed and accuracy.To address these challenges,this dissertation proposes a flow-based video object detection framework,that uses feature flow to model the relative motion between adja-cent frames at the feature level,and automatically extracts the required temporary infor-mation through end-to-end learning.The proposed framework can achieve high-speed and high-accuracy video object detection.The main contributions and innovations are summarized as follows:·A high-speed video object detector based on feature propagation is proposed.The proposed method only applies the expensive feature extraction sub-network on sparse key frames.The feature maps of other non-key frames are propagated from key frames via flow-guided feature propagation.After propagation,the detection sub-network is applied on these propagated feature maps to generate detection results.Compared with the per-frame image object detection,the pro-posed method can effectively improve the detection speed while maintaining the detection accuracy.·A high-accuracy video object detector based on feature aggregation is proposed.The proposed method enhances the feature quality by aggregating nearby fea-ture maps along the motion paths.After aggregation,the detection sub-network is applied on these aggregated feature maps to generate detection results.The proposed method does not require manually designed rules for multi-frame ag-gregation and can effectively improve the detection accuracy.·A high-speed and high-accuracy video object detector based on recursive feature aggregation and adaptive feature extraction is proposed.The proposed method only applies feature extraction and recursive feature aggregation on sparse key frames,and adaptively performs partial feature extraction and key frame selec-tion based on the estimated quality of propagated feature.The proposed method can achieve high accuracy while meeting real-time speed requirements,and has been verified on both high-performance GPU server and resource-limited mobile device.